fs/ceph/file.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 #include <linux/ceph/ceph_debug.h>
   3 #include <linux/ceph/striper.h>
   4
   5 #include <linux/module.h>
   6 #include <linux/sched.h>
   7 #include <linux/slab.h>
   8 #include <linux/file.h>
   9 #include <linux/mount.h>
  10 #include <linux/namei.h>
  11 #include <linux/writeback.h>
  12 #include <linux/falloc.h>
  13 #include <linux/iversion.h>
  14
  15 #include "super.h"
  16 #include "mds_client.h"
  17 #include "cache.h"
  18
  19 static __le32 ceph_flags_sys2wire(u32 flags)
  20 {
  21         u32 wire_flags = 0;
  22
  23         switch (flags & O_ACCMODE) {
  24         case O_RDONLY:
  25                 wire_flags |= CEPH_O_RDONLY;
  26                 break;
  27         case O_WRONLY:
  28                 wire_flags |= CEPH_O_WRONLY;
  29                 break;
  30         case O_RDWR:
  31                 wire_flags |= CEPH_O_RDWR;
  32                 break;
  33         }
  34
  35         flags &= ~O_ACCMODE;
  36
  37 #define ceph_sys2wire(a) if (flags & a) { wire_flags |= CEPH_##a; flags &= ~a; }
  38
  39         ceph_sys2wire(O_CREAT);
  40         ceph_sys2wire(O_EXCL);
  41         ceph_sys2wire(O_TRUNC);
  42         ceph_sys2wire(O_DIRECTORY);
  43         ceph_sys2wire(O_NOFOLLOW);
  44
  45 #undef ceph_sys2wire
  46
  47         if (flags)
  48                 dout("unused open flags: %x\n", flags);
  49
  50         return cpu_to_le32(wire_flags);
  51 }
  52
  53 /*
  54  * Ceph file operations
  55  *
  56  * Implement basic open/close functionality, and implement
  57  * read/write.
  58  *
  59  * We implement three modes of file I/O:
  60  *  - buffered uses the generic_file_aio_{read,write} helpers
  61  *
  62  *  - synchronous is used when there is multi-client read/write
  63  *    sharing, avoids the page cache, and synchronously waits for an
  64  *    ack from the OSD.
  65  *
  66  *  - direct io takes the variant of the sync path that references
  67  *    user pages directly.
  68  *
  69  * fsync() flushes and waits on dirty pages, but just queues metadata
  70  * for writeback: since the MDS can recover size and mtime there is no
  71  * need to wait for MDS acknowledgement.
  72  */
  73
  74 /*
  75  * How many pages to get in one call to iov_iter_get_pages().  This
  76  * determines the size of the on-stack array used as a buffer.
  77  */
  78 #define ITER_GET_BVECS_PAGES    64
  79
  80 static ssize_t __iter_get_bvecs(struct iov_iter *iter, size_t maxsize,
  81                                 struct bio_vec *bvecs)
  82 {
  83         size_t size = 0;
  84         int bvec_idx = 0;
  85
  86         if (maxsize > iov_iter_count(iter))
  87                 maxsize = iov_iter_count(iter);
  88
  89         while (size < maxsize) {
  90                 struct page *pages[ITER_GET_BVECS_PAGES];
  91                 ssize_t bytes;
  92                 size_t start;
  93                 int idx = 0;
  94
  95                 bytes = iov_iter_get_pages(iter, pages, maxsize - size,
  96                                            ITER_GET_BVECS_PAGES, &start);
  97                 if (bytes < 0)
  98                         return size ?: bytes;
  99
 100                 iov_iter_advance(iter, bytes);
 101                 size += bytes;
 102
 103                 for ( ; bytes; idx++, bvec_idx++) {
 104                         struct bio_vec bv = {
 105                                 .bv_page = pages[idx],
 106                                 .bv_len = min_t(int, bytes, PAGE_SIZE - start),
 107                                 .bv_offset = start,
 108                         };
 109
 110                         bvecs[bvec_idx] = bv;
 111                         bytes -= bv.bv_len;
 112                         start = 0;
 113                 }
 114         }
 115
 116         return size;
 117 }
 118
 119 /*
 120  * iov_iter_get_pages() only considers one iov_iter segment, no matter
 121  * what maxsize or maxpages are given.  For ITER_BVEC that is a single
 122  * page.
 123  *
 124  * Attempt to get up to @maxsize bytes worth of pages from @iter.
 125  * Return the number of bytes in the created bio_vec array, or an error.
 126  */
 127 static ssize_t iter_get_bvecs_alloc(struct iov_iter *iter, size_t maxsize,
 128                                     struct bio_vec **bvecs, int *num_bvecs)
 129 {
 130         struct bio_vec *bv;
 131         size_t orig_count = iov_iter_count(iter);
 132         ssize_t bytes;
 133         int npages;
 134
 135         iov_iter_truncate(iter, maxsize);
 136         npages = iov_iter_npages(iter, INT_MAX);
 137         iov_iter_reexpand(iter, orig_count);
 138
 139         /*
 140          * __iter_get_bvecs() may populate only part of the array -- zero it
 141          * out.
 142          */
 143         bv = kvmalloc_array(npages, sizeof(*bv), GFP_KERNEL | __GFP_ZERO);
 144         if (!bv)
 145                 return -ENOMEM;
 146
 147         bytes = __iter_get_bvecs(iter, maxsize, bv);
 148         if (bytes < 0) {
 149                 /*
 150                  * No pages were pinned -- just free the array.
 151                  */
 152                 kvfree(bv);
 153                 return bytes;
 154         }
 155
 156         *bvecs = bv;
 157         *num_bvecs = npages;
 158         return bytes;
 159 }
 160
 161 static void put_bvecs(struct bio_vec *bvecs, int num_bvecs, bool should_dirty)
 162 {
 163         int i;
 164
 165         for (i = 0; i < num_bvecs; i++) {
 166                 if (bvecs[i].bv_page) {
 167                         if (should_dirty)
 168                                 set_page_dirty_lock(bvecs[i].bv_page);
 169                         put_page(bvecs[i].bv_page);
 170                 }
 171         }
 172         kvfree(bvecs);
 173 }
 174
 175 /*
 176  * Prepare an open request.  Preallocate ceph_cap to avoid an
 177  * inopportune ENOMEM later.
 178  */
 179 static struct ceph_mds_request *
 180 prepare_open_request(struct super_block *sb, int flags, int create_mode)
 181 {
 182         struct ceph_fs_client *fsc = ceph_sb_to_client(sb);
 183         struct ceph_mds_client *mdsc = fsc->mdsc;
 184         struct ceph_mds_request *req;
 185         int want_auth = USE_ANY_MDS;
 186         int op = (flags & O_CREAT) ? CEPH_MDS_OP_CREATE : CEPH_MDS_OP_OPEN;
 187
 188         if (flags & (O_WRONLY|O_RDWR|O_CREAT|O_TRUNC))
 189                 want_auth = USE_AUTH_MDS;
 190
 191         req = ceph_mdsc_create_request(mdsc, op, want_auth);
 192         if (IS_ERR(req))
 193                 goto out;
 194         req->r_fmode = ceph_flags_to_mode(flags);
 195         req->r_args.open.flags = ceph_flags_sys2wire(flags);
 196         req->r_args.open.mode = cpu_to_le32(create_mode);
 197 out:
 198         return req;
 199 }
 200
 201 static int ceph_init_file_info(struct inode *inode, struct file *file,
 202                                         int fmode, bool isdir)
 203 {
 204         struct ceph_file_info *fi;
 205
 206         dout("%s %p %p 0%o (%s)\n", __func__, inode, file,
 207                         inode->i_mode, isdir ? "dir" : "regular");
 208         BUG_ON(inode->i_fop->release != ceph_release);
 209
 210         if (isdir) {
 211                 struct ceph_dir_file_info *dfi =
 212                         kmem_cache_zalloc(ceph_dir_file_cachep, GFP_KERNEL);
 213                 if (!dfi) {
 214                         ceph_put_fmode(ceph_inode(inode), fmode); /* clean up */
 215                         return -ENOMEM;
 216                 }
 217
 218                 file->private_data = dfi;
 219                 fi = &dfi->file_info;
 220                 dfi->next_offset = 2;
 221                 dfi->readdir_cache_idx = -1;
 222         } else {
 223                 fi = kmem_cache_zalloc(ceph_file_cachep, GFP_KERNEL);
 224                 if (!fi) {
 225                         ceph_put_fmode(ceph_inode(inode), fmode); /* clean up */
 226                         return -ENOMEM;
 227                 }
 228
 229                 file->private_data = fi;
 230         }
 231
 232         fi->fmode = fmode;
 233         spin_lock_init(&fi->rw_contexts_lock);
 234         INIT_LIST_HEAD(&fi->rw_contexts);
 235
 236         return 0;
 237 }
 238
 239 /*
 240  * initialize private struct file data.
 241  * if we fail, clean up by dropping fmode reference on the ceph_inode
 242  */
 243 static int ceph_init_file(struct inode *inode, struct file *file, int fmode)
 244 {
 245         int ret = 0;
 246
 247         switch (inode->i_mode & S_IFMT) {
 248         case S_IFREG:
 249                 ceph_fscache_register_inode_cookie(inode);
 250                 ceph_fscache_file_set_cookie(inode, file);
 251                 /* fall through */
 252         case S_IFDIR:
 253                 ret = ceph_init_file_info(inode, file, fmode,
 254                                                 S_ISDIR(inode->i_mode));
 255                 if (ret)
 256                         return ret;
 257                 break;
 258
 259         case S_IFLNK:
 260                 dout("init_file %p %p 0%o (symlink)\n", inode, file,
 261                      inode->i_mode);
 262                 ceph_put_fmode(ceph_inode(inode), fmode); /* clean up */
 263                 break;
 264
 265         default:
 266                 dout("init_file %p %p 0%o (special)\n", inode, file,
 267                      inode->i_mode);
 268                 /*
 269                  * we need to drop the open ref now, since we don't
 270                  * have .release set to ceph_release.
 271                  */
 272                 ceph_put_fmode(ceph_inode(inode), fmode); /* clean up */
 273                 BUG_ON(inode->i_fop->release == ceph_release);
 274
 275                 /* call the proper open fop */
 276                 ret = inode->i_fop->open(inode, file);
 277         }
 278         return ret;
 279 }
 280
 281 /*
 282  * try renew caps after session gets killed.
 283  */
 284 int ceph_renew_caps(struct inode *inode)
 285 {
 286         struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
 287         struct ceph_inode_info *ci = ceph_inode(inode);
 288         struct ceph_mds_request *req;
 289         int err, flags, wanted;
 290
 291         spin_lock(&ci->i_ceph_lock);
 292         wanted = __ceph_caps_file_wanted(ci);
 293         if (__ceph_is_any_real_caps(ci) &&
 294             (!(wanted & CEPH_CAP_ANY_WR) || ci->i_auth_cap)) {
 295                 int issued = __ceph_caps_issued(ci, NULL);
 296                 spin_unlock(&ci->i_ceph_lock);
 297                 dout("renew caps %p want %s issued %s updating mds_wanted\n",
 298                      inode, ceph_cap_string(wanted), ceph_cap_string(issued));
 299                 ceph_check_caps(ci, 0, NULL);
 300                 return 0;
 301         }
 302         spin_unlock(&ci->i_ceph_lock);
 303
 304         flags = 0;
 305         if ((wanted & CEPH_CAP_FILE_RD) && (wanted & CEPH_CAP_FILE_WR))
 306                 flags = O_RDWR;
 307         else if (wanted & CEPH_CAP_FILE_RD)
 308                 flags = O_RDONLY;
 309         else if (wanted & CEPH_CAP_FILE_WR)
 310                 flags = O_WRONLY;
 311 #ifdef O_LAZY
 312         if (wanted & CEPH_CAP_FILE_LAZYIO)
 313                 flags |= O_LAZY;
 314 #endif
 315
 316         req = prepare_open_request(inode->i_sb, flags, 0);
 317         if (IS_ERR(req)) {
 318                 err = PTR_ERR(req);
 319                 goto out;
 320         }
 321
 322         req->r_inode = inode;
 323         ihold(inode);
 324         req->r_num_caps = 1;
 325         req->r_fmode = -1;
 326
 327         err = ceph_mdsc_do_request(mdsc, NULL, req);
 328         ceph_mdsc_put_request(req);
 329 out:
 330         dout("renew caps %p open result=%d\n", inode, err);
 331         return err < 0 ? err : 0;
 332 }
 333
 334 /*
 335  * If we already have the requisite capabilities, we can satisfy
 336  * the open request locally (no need to request new caps from the
 337  * MDS).  We do, however, need to inform the MDS (asynchronously)
 338  * if our wanted caps set expands.
 339  */
 340 int ceph_open(struct inode *inode, struct file *file)
 341 {
 342         struct ceph_inode_info *ci = ceph_inode(inode);
 343         struct ceph_fs_client *fsc = ceph_sb_to_client(inode->i_sb);
 344         struct ceph_mds_client *mdsc = fsc->mdsc;
 345         struct ceph_mds_request *req;
 346         struct ceph_file_info *fi = file->private_data;
 347         int err;
 348         int flags, fmode, wanted;
 349
 350         if (fi) {
 351                 dout("open file %p is already opened\n", file);
 352                 return 0;
 353         }
 354
 355         /* filter out O_CREAT|O_EXCL; vfs did that already.  yuck. */
 356         flags = file->f_flags & ~(O_CREAT|O_EXCL);
 357         if (S_ISDIR(inode->i_mode))
 358                 flags = O_DIRECTORY;  /* mds likes to know */
 359
 360         dout("open inode %p ino %llx.%llx file %p flags %d (%d)\n", inode,
 361              ceph_vinop(inode), file, flags, file->f_flags);
 362         fmode = ceph_flags_to_mode(flags);
 363         wanted = ceph_caps_for_mode(fmode);
 364
 365         /* snapped files are read-only */
 366         if (ceph_snap(inode) != CEPH_NOSNAP && (file->f_mode & FMODE_WRITE))
 367                 return -EROFS;
 368
 369         /* trivially open snapdir */
 370         if (ceph_snap(inode) == CEPH_SNAPDIR) {
 371                 spin_lock(&ci->i_ceph_lock);
 372                 __ceph_get_fmode(ci, fmode);
 373                 spin_unlock(&ci->i_ceph_lock);
 374                 return ceph_init_file(inode, file, fmode);
 375         }
 376
 377         /*
 378          * No need to block if we have caps on the auth MDS (for
 379          * write) or any MDS (for read).  Update wanted set
 380          * asynchronously.
 381          */
 382         spin_lock(&ci->i_ceph_lock);
 383         if (__ceph_is_any_real_caps(ci) &&
 384             (((fmode & CEPH_FILE_MODE_WR) == 0) || ci->i_auth_cap)) {
 385                 int mds_wanted = __ceph_caps_mds_wanted(ci, true);
 386                 int issued = __ceph_caps_issued(ci, NULL);
 387
 388                 dout("open %p fmode %d want %s issued %s using existing\n",
 389                      inode, fmode, ceph_cap_string(wanted),
 390                      ceph_cap_string(issued));
 391                 __ceph_get_fmode(ci, fmode);
 392                 spin_unlock(&ci->i_ceph_lock);
 393
 394                 /* adjust wanted? */
 395                 if ((issued & wanted) != wanted &&
 396                     (mds_wanted & wanted) != wanted &&
 397                     ceph_snap(inode) != CEPH_SNAPDIR)
 398                         ceph_check_caps(ci, 0, NULL);
 399
 400                 return ceph_init_file(inode, file, fmode);
 401         } else if (ceph_snap(inode) != CEPH_NOSNAP &&
 402                    (ci->i_snap_caps & wanted) == wanted) {
 403                 __ceph_get_fmode(ci, fmode);
 404                 spin_unlock(&ci->i_ceph_lock);
 405                 return ceph_init_file(inode, file, fmode);
 406         }
 407
 408         spin_unlock(&ci->i_ceph_lock);
 409
 410         dout("open fmode %d wants %s\n", fmode, ceph_cap_string(wanted));
 411         req = prepare_open_request(inode->i_sb, flags, 0);
 412         if (IS_ERR(req)) {
 413                 err = PTR_ERR(req);
 414                 goto out;
 415         }
 416         req->r_inode = inode;
 417         ihold(inode);
 418
 419         req->r_num_caps = 1;
 420         err = ceph_mdsc_do_request(mdsc, NULL, req);
 421         if (!err)
 422                 err = ceph_init_file(inode, file, req->r_fmode);
 423         ceph_mdsc_put_request(req);
 424         dout("open result=%d on %llx.%llx\n", err, ceph_vinop(inode));
 425 out:
 426         return err;
 427 }
 428
 429
 430 /*
 431  * Do a lookup + open with a single request.  If we get a non-existent
 432  * file or symlink, return 1 so the VFS can retry.
 433  */
 434 int ceph_atomic_open(struct inode *dir, struct dentry *dentry,
 435                      struct file *file, unsigned flags, umode_t mode)
 436 {
 437         struct ceph_fs_client *fsc = ceph_sb_to_client(dir->i_sb);
 438         struct ceph_mds_client *mdsc = fsc->mdsc;
 439         struct ceph_mds_request *req;
 440         struct dentry *dn;
 441         struct ceph_acl_sec_ctx as_ctx = {};
 442         int mask;
 443         int err;
 444
 445         dout("atomic_open %p dentry %p '%pd' %s flags %d mode 0%o\n",
 446              dir, dentry, dentry,
 447              d_unhashed(dentry) ? "unhashed" : "hashed", flags, mode);
 448
 449         if (dentry->d_name.len > NAME_MAX)
 450                 return -ENAMETOOLONG;
 451
 452         if (flags & O_CREAT) {
 453                 if (ceph_quota_is_max_files_exceeded(dir))
 454                         return -EDQUOT;
 455                 err = ceph_pre_init_acls(dir, &mode, &as_ctx);
 456                 if (err < 0)
 457                         return err;
 458                 err = ceph_security_init_secctx(dentry, mode, &as_ctx);
 459                 if (err < 0)
 460                         goto out_ctx;
 461         } else if (!d_in_lookup(dentry)) {
 462                 /* If it's not being looked up, it's negative */
 463                 return -ENOENT;
 464         }
 465
 466         /* do the open */
 467         req = prepare_open_request(dir->i_sb, flags, mode);
 468         if (IS_ERR(req)) {
 469                 err = PTR_ERR(req);
 470                 goto out_ctx;
 471         }
 472         req->r_dentry = dget(dentry);
 473         req->r_num_caps = 2;
 474         if (flags & O_CREAT) {
 475                 req->r_dentry_drop = CEPH_CAP_FILE_SHARED | CEPH_CAP_AUTH_EXCL;
 476                 req->r_dentry_unless = CEPH_CAP_FILE_EXCL;
 477                 if (as_ctx.pagelist) {
 478                         req->r_pagelist = as_ctx.pagelist;
 479                         as_ctx.pagelist = NULL;
 480                 }
 481         }
 482
 483        mask = CEPH_STAT_CAP_INODE | CEPH_CAP_AUTH_SHARED;
 484        if (ceph_security_xattr_wanted(dir))
 485                mask |= CEPH_CAP_XATTR_SHARED;
 486        req->r_args.open.mask = cpu_to_le32(mask);
 487
 488         req->r_parent = dir;
 489         set_bit(CEPH_MDS_R_PARENT_LOCKED, &req->r_req_flags);
 490         err = ceph_mdsc_do_request(mdsc,
 491                                    (flags & (O_CREAT|O_TRUNC)) ? dir : NULL,
 492                                    req);
 493         err = ceph_handle_snapdir(req, dentry, err);
 494         if (err)
 495                 goto out_req;
 496
 497         if ((flags & O_CREAT) && !req->r_reply_info.head->is_dentry)
 498                 err = ceph_handle_notrace_create(dir, dentry);
 499
 500         if (d_in_lookup(dentry)) {
 501                 dn = ceph_finish_lookup(req, dentry, err);
 502                 if (IS_ERR(dn))
 503                         err = PTR_ERR(dn);
 504         } else {
 505                 /* we were given a hashed negative dentry */
 506                 dn = NULL;
 507         }
 508         if (err)
 509                 goto out_req;
 510         if (dn || d_really_is_negative(dentry) || d_is_symlink(dentry)) {
 511                 /* make vfs retry on splice, ENOENT, or symlink */
 512                 dout("atomic_open finish_no_open on dn %p\n", dn);
 513                 err = finish_no_open(file, dn);
 514         } else {
 515                 dout("atomic_open finish_open on dn %p\n", dn);
 516                 if (req->r_op == CEPH_MDS_OP_CREATE && req->r_reply_info.has_create_ino) {
 517                         ceph_init_inode_acls(d_inode(dentry), &as_ctx);
 518                         file->f_mode |= FMODE_CREATED;
 519                 }
 520                 err = finish_open(file, dentry, ceph_open);
 521         }
 522 out_req:
 523         if (!req->r_err && req->r_target_inode)
 524                 ceph_put_fmode(ceph_inode(req->r_target_inode), req->r_fmode);
 525         ceph_mdsc_put_request(req);
 526 out_ctx:
 527         ceph_release_acl_sec_ctx(&as_ctx);
 528         dout("atomic_open result=%d\n", err);
 529         return err;
 530 }
 531
 532 int ceph_release(struct inode *inode, struct file *file)
 533 {
 534         struct ceph_inode_info *ci = ceph_inode(inode);
 535
 536         if (S_ISDIR(inode->i_mode)) {
 537                 struct ceph_dir_file_info *dfi = file->private_data;
 538                 dout("release inode %p dir file %p\n", inode, file);
 539                 WARN_ON(!list_empty(&dfi->file_info.rw_contexts));
 540
 541                 ceph_put_fmode(ci, dfi->file_info.fmode);
 542
 543                 if (dfi->last_readdir)
 544                         ceph_mdsc_put_request(dfi->last_readdir);
 545                 kfree(dfi->last_name);
 546                 kfree(dfi->dir_info);
 547                 kmem_cache_free(ceph_dir_file_cachep, dfi);
 548         } else {
 549                 struct ceph_file_info *fi = file->private_data;
 550                 dout("release inode %p regular file %p\n", inode, file);
 551                 WARN_ON(!list_empty(&fi->rw_contexts));
 552
 553                 ceph_put_fmode(ci, fi->fmode);
 554                 kmem_cache_free(ceph_file_cachep, fi);
 555         }
 556
 557         /* wake up anyone waiting for caps on this inode */
 558         wake_up_all(&ci->i_cap_wq);
 559         return 0;
 560 }
 561
 562 enum {
 563         HAVE_RETRIED = 1,
 564         CHECK_EOF =    2,
 565         READ_INLINE =  3,
 566 };
 567
 568 /*
 569  * Completely synchronous read and write methods.  Direct from __user
 570  * buffer to osd, or directly to user pages (if O_DIRECT).
 571  *
 572  * If the read spans object boundary, just do multiple reads.  (That's not
 573  * atomic, but good enough for now.)
 574  *
 575  * If we get a short result from the OSD, check against i_size; we need to
 576  * only return a short read to the caller if we hit EOF.
 577  */
 578 static ssize_t ceph_sync_read(struct kiocb *iocb, struct iov_iter *to,
 579                               int *retry_op)
 580 {
 581         struct file *file = iocb->ki_filp;
 582         struct inode *inode = file_inode(file);
 583         struct ceph_inode_info *ci = ceph_inode(inode);
 584         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
 585         struct ceph_osd_client *osdc = &fsc->client->osdc;
 586         ssize_t ret;
 587         u64 off = iocb->ki_pos;
 588         u64 len = iov_iter_count(to);
 589
 590         dout("sync_read on file %p %llu~%u %s\n", file, off, (unsigned)len,
 591              (file->f_flags & O_DIRECT) ? "O_DIRECT" : "");
 592
 593         if (!len)
 594                 return 0;
 595         /*
 596          * flush any page cache pages in this range.  this
 597          * will make concurrent normal and sync io slow,
 598          * but it will at least behave sensibly when they are
 599          * in sequence.
 600          */
 601         ret = filemap_write_and_wait_range(inode->i_mapping,
 602                                            off, off + len - 1);
 603         if (ret < 0)
 604                 return ret;
 605
 606         ret = 0;
 607         while ((len = iov_iter_count(to)) > 0) {
 608                 struct ceph_osd_request *req;
 609                 struct page **pages;
 610                 int num_pages;
 611                 size_t page_off;
 612                 u64 i_size;
 613                 bool more;
 614
 615                 req = ceph_osdc_new_request(osdc, &ci->i_layout,
 616                                         ci->i_vino, off, &len, 0, 1,
 617                                         CEPH_OSD_OP_READ, CEPH_OSD_FLAG_READ,
 618                                         NULL, ci->i_truncate_seq,
 619                                         ci->i_truncate_size, false);
 620                 if (IS_ERR(req)) {
 621                         ret = PTR_ERR(req);
 622                         break;
 623                 }
 624
 625                 more = len < iov_iter_count(to);
 626
 627                 if (unlikely(iov_iter_is_pipe(to))) {
 628                         ret = iov_iter_get_pages_alloc(to, &pages, len,
 629                                                        &page_off);
 630                         if (ret <= 0) {
 631                                 ceph_osdc_put_request(req);
 632                                 ret = -ENOMEM;
 633                                 break;
 634                         }
 635                         num_pages = DIV_ROUND_UP(ret + page_off, PAGE_SIZE);
 636                         if (ret < len) {
 637                                 len = ret;
 638                                 osd_req_op_extent_update(req, 0, len);
 639                                 more = false;
 640                         }
 641                 } else {
 642                         num_pages = calc_pages_for(off, len);
 643                         page_off = off & ~PAGE_MASK;
 644                         pages = ceph_alloc_page_vector(num_pages, GFP_KERNEL);
 645                         if (IS_ERR(pages)) {
 646                                 ceph_osdc_put_request(req);
 647                                 ret = PTR_ERR(pages);
 648                                 break;
 649                         }
 650                 }
 651
 652                 osd_req_op_extent_osd_data_pages(req, 0, pages, len, page_off,
 653                                                  false, false);
 654                 ret = ceph_osdc_start_request(osdc, req, false);
 655                 if (!ret)
 656                         ret = ceph_osdc_wait_request(osdc, req);
 657                 ceph_osdc_put_request(req);
 658
 659                 i_size = i_size_read(inode);
 660                 dout("sync_read %llu~%llu got %zd i_size %llu%s\n",
 661                      off, len, ret, i_size, (more ? " MORE" : ""));
 662
 663                 if (ret == -ENOENT)
 664                         ret = 0;
 665                 if (ret >= 0 && ret < len && (off + ret < i_size)) {
 666                         int zlen = min(len - ret, i_size - off - ret);
 667                         int zoff = page_off + ret;
 668                         dout("sync_read zero gap %llu~%llu\n",
 669                              off + ret, off + ret + zlen);
 670                         ceph_zero_page_vector_range(zoff, zlen, pages);
 671                         ret += zlen;
 672                 }
 673
 674                 if (unlikely(iov_iter_is_pipe(to))) {
 675                         if (ret > 0) {
 676                                 iov_iter_advance(to, ret);
 677                                 off += ret;
 678                         } else {
 679                                 iov_iter_advance(to, 0);
 680                         }
 681                         ceph_put_page_vector(pages, num_pages, false);
 682                 } else {
 683                         int idx = 0;
 684                         size_t left = ret > 0 ? ret : 0;
 685                         while (left > 0) {
 686                                 size_t len, copied;
 687                                 page_off = off & ~PAGE_MASK;
 688                                 len = min_t(size_t, left, PAGE_SIZE - page_off);
 689                                 copied = copy_page_to_iter(pages[idx++],
 690                                                            page_off, len, to);
 691                                 off += copied;
 692                                 left -= copied;
 693                                 if (copied < len) {
 694                                         ret = -EFAULT;
 695                                         break;
 696                                 }
 697                         }
 698                         ceph_release_page_vector(pages, num_pages);
 699                 }
 700
 701                 if (ret <= 0 || off >= i_size || !more)
 702                         break;
 703         }
 704
 705         if (off > iocb->ki_pos) {
 706                 if (ret >= 0 &&
 707                     iov_iter_count(to) > 0 && off >= i_size_read(inode))
 708                         *retry_op = CHECK_EOF;
 709                 ret = off - iocb->ki_pos;
 710                 iocb->ki_pos = off;
 711         }
 712
 713         dout("sync_read result %zd retry_op %d\n", ret, *retry_op);
 714         return ret;
 715 }
 716
 717 struct ceph_aio_request {
 718         struct kiocb *iocb;
 719         size_t total_len;
 720         bool write;
 721         bool should_dirty;
 722         int error;
 723         struct list_head osd_reqs;
 724         unsigned num_reqs;
 725         atomic_t pending_reqs;
 726         struct timespec64 mtime;
 727         struct ceph_cap_flush *prealloc_cf;
 728 };
 729
 730 struct ceph_aio_work {
 731         struct work_struct work;
 732         struct ceph_osd_request *req;
 733 };
 734
 735 static void ceph_aio_retry_work(struct work_struct *work);
 736
 737 static void ceph_aio_complete(struct inode *inode,
 738                               struct ceph_aio_request *aio_req)
 739 {
 740         struct ceph_inode_info *ci = ceph_inode(inode);
 741         int ret;
 742
 743         if (!atomic_dec_and_test(&aio_req->pending_reqs))
 744                 return;
 745
 746         ret = aio_req->error;
 747         if (!ret)
 748                 ret = aio_req->total_len;
 749
 750         dout("ceph_aio_complete %p rc %d\n", inode, ret);
 751
 752         if (ret >= 0 && aio_req->write) {
 753                 int dirty;
 754
 755                 loff_t endoff = aio_req->iocb->ki_pos + aio_req->total_len;
 756                 if (endoff > i_size_read(inode)) {
 757                         if (ceph_inode_set_size(inode, endoff))
 758                                 ceph_check_caps(ci, CHECK_CAPS_AUTHONLY, NULL);
 759                 }
 760
 761                 spin_lock(&ci->i_ceph_lock);
 762                 ci->i_inline_version = CEPH_INLINE_NONE;
 763                 dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR,
 764                                                &aio_req->prealloc_cf);
 765                 spin_unlock(&ci->i_ceph_lock);
 766                 if (dirty)
 767                         __mark_inode_dirty(inode, dirty);
 768
 769         }
 770
 771         ceph_put_cap_refs(ci, (aio_req->write ? CEPH_CAP_FILE_WR :
 772                                                 CEPH_CAP_FILE_RD));
 773
 774         aio_req->iocb->ki_complete(aio_req->iocb, ret, 0);
 775
 776         ceph_free_cap_flush(aio_req->prealloc_cf);
 777         kfree(aio_req);
 778 }
 779
 780 static void ceph_aio_complete_req(struct ceph_osd_request *req)
 781 {
 782         int rc = req->r_result;
 783         struct inode *inode = req->r_inode;
 784         struct ceph_aio_request *aio_req = req->r_priv;
 785         struct ceph_osd_data *osd_data = osd_req_op_extent_osd_data(req, 0);
 786
 787         BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_BVECS);
 788         BUG_ON(!osd_data->num_bvecs);
 789
 790         dout("ceph_aio_complete_req %p rc %d bytes %u\n",
 791              inode, rc, osd_data->bvec_pos.iter.bi_size);
 792
 793         if (rc == -EOLDSNAPC) {
 794                 struct ceph_aio_work *aio_work;
 795                 BUG_ON(!aio_req->write);
 796
 797                 aio_work = kmalloc(sizeof(*aio_work), GFP_NOFS);
 798                 if (aio_work) {
 799                         INIT_WORK(&aio_work->work, ceph_aio_retry_work);
 800                         aio_work->req = req;
 801                         queue_work(ceph_inode_to_client(inode)->inode_wq,
 802                                    &aio_work->work);
 803                         return;
 804                 }
 805                 rc = -ENOMEM;
 806         } else if (!aio_req->write) {
 807                 if (rc == -ENOENT)
 808                         rc = 0;
 809                 if (rc >= 0 && osd_data->bvec_pos.iter.bi_size > rc) {
 810                         struct iov_iter i;
 811                         int zlen = osd_data->bvec_pos.iter.bi_size - rc;
 812
 813                         /*
 814                          * If read is satisfied by single OSD request,
 815                          * it can pass EOF. Otherwise read is within
 816                          * i_size.
 817                          */
 818                         if (aio_req->num_reqs == 1) {
 819                                 loff_t i_size = i_size_read(inode);
 820                                 loff_t endoff = aio_req->iocb->ki_pos + rc;
 821                                 if (endoff < i_size)
 822                                         zlen = min_t(size_t, zlen,
 823                                                      i_size - endoff);
 824                                 aio_req->total_len = rc + zlen;
 825                         }
 826
 827                         iov_iter_bvec(&i, READ, osd_data->bvec_pos.bvecs,
 828                                       osd_data->num_bvecs,
 829                                       osd_data->bvec_pos.iter.bi_size);
 830                         iov_iter_advance(&i, rc);
 831                         iov_iter_zero(zlen, &i);
 832                 }
 833         }
 834
 835         put_bvecs(osd_data->bvec_pos.bvecs, osd_data->num_bvecs,
 836                   aio_req->should_dirty);
 837         ceph_osdc_put_request(req);
 838
 839         if (rc < 0)
 840                 cmpxchg(&aio_req->error, 0, rc);
 841
 842         ceph_aio_complete(inode, aio_req);
 843         return;
 844 }
 845
 846 static void ceph_aio_retry_work(struct work_struct *work)
 847 {
 848         struct ceph_aio_work *aio_work =
 849                 container_of(work, struct ceph_aio_work, work);
 850         struct ceph_osd_request *orig_req = aio_work->req;
 851         struct ceph_aio_request *aio_req = orig_req->r_priv;
 852         struct inode *inode = orig_req->r_inode;
 853         struct ceph_inode_info *ci = ceph_inode(inode);
 854         struct ceph_snap_context *snapc;
 855         struct ceph_osd_request *req;
 856         int ret;
 857
 858         spin_lock(&ci->i_ceph_lock);
 859         if (__ceph_have_pending_cap_snap(ci)) {
 860                 struct ceph_cap_snap *capsnap =
 861                         list_last_entry(&ci->i_cap_snaps,
 862                                         struct ceph_cap_snap,
 863                                         ci_item);
 864                 snapc = ceph_get_snap_context(capsnap->context);
 865         } else {
 866                 BUG_ON(!ci->i_head_snapc);
 867                 snapc = ceph_get_snap_context(ci->i_head_snapc);
 868         }
 869         spin_unlock(&ci->i_ceph_lock);
 870
 871         req = ceph_osdc_alloc_request(orig_req->r_osdc, snapc, 1,
 872                         false, GFP_NOFS);
 873         if (!req) {
 874                 ret = -ENOMEM;
 875                 req = orig_req;
 876                 goto out;
 877         }
 878
 879         req->r_flags = /* CEPH_OSD_FLAG_ORDERSNAP | */ CEPH_OSD_FLAG_WRITE;
 880         ceph_oloc_copy(&req->r_base_oloc, &orig_req->r_base_oloc);
 881         ceph_oid_copy(&req->r_base_oid, &orig_req->r_base_oid);
 882
 883         req->r_ops[0] = orig_req->r_ops[0];
 884
 885         req->r_mtime = aio_req->mtime;
 886         req->r_data_offset = req->r_ops[0].extent.offset;
 887
 888         ret = ceph_osdc_alloc_messages(req, GFP_NOFS);
 889         if (ret) {
 890                 ceph_osdc_put_request(req);
 891                 req = orig_req;
 892                 goto out;
 893         }
 894
 895         ceph_osdc_put_request(orig_req);
 896
 897         req->r_callback = ceph_aio_complete_req;
 898         req->r_inode = inode;
 899         req->r_priv = aio_req;
 900
 901         ret = ceph_osdc_start_request(req->r_osdc, req, false);
 902 out:
 903         if (ret < 0) {
 904                 req->r_result = ret;
 905                 ceph_aio_complete_req(req);
 906         }
 907
 908         ceph_put_snap_context(snapc);
 909         kfree(aio_work);
 910 }
 911
 912 static ssize_t
 913 ceph_direct_read_write(struct kiocb *iocb, struct iov_iter *iter,
 914                        struct ceph_snap_context *snapc,
 915                        struct ceph_cap_flush **pcf)
 916 {
 917         struct file *file = iocb->ki_filp;
 918         struct inode *inode = file_inode(file);
 919         struct ceph_inode_info *ci = ceph_inode(inode);
 920         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
 921         struct ceph_vino vino;
 922         struct ceph_osd_request *req;
 923         struct bio_vec *bvecs;
 924         struct ceph_aio_request *aio_req = NULL;
 925         int num_pages = 0;
 926         int flags;
 927         int ret;
 928         struct timespec64 mtime = current_time(inode);
 929         size_t count = iov_iter_count(iter);
 930         loff_t pos = iocb->ki_pos;
 931         bool write = iov_iter_rw(iter) == WRITE;
 932         bool should_dirty = !write && iter_is_iovec(iter);
 933
 934         if (write && ceph_snap(file_inode(file)) != CEPH_NOSNAP)
 935                 return -EROFS;
 936
 937         dout("sync_direct_%s on file %p %lld~%u snapc %p seq %lld\n",
 938              (write ? "write" : "read"), file, pos, (unsigned)count,
 939              snapc, snapc ? snapc->seq : 0);
 940
 941         ret = filemap_write_and_wait_range(inode->i_mapping,
 942                                            pos, pos + count - 1);
 943         if (ret < 0)
 944                 return ret;
 945
 946         if (write) {
 947                 int ret2 = invalidate_inode_pages2_range(inode->i_mapping,
 948                                         pos >> PAGE_SHIFT,
 949                                         (pos + count - 1) >> PAGE_SHIFT);
 950                 if (ret2 < 0)
 951                         dout("invalidate_inode_pages2_range returned %d\n", ret2);
 952
 953                 flags = /* CEPH_OSD_FLAG_ORDERSNAP | */ CEPH_OSD_FLAG_WRITE;
 954         } else {
 955                 flags = CEPH_OSD_FLAG_READ;
 956         }
 957
 958         while (iov_iter_count(iter) > 0) {
 959                 u64 size = iov_iter_count(iter);
 960                 ssize_t len;
 961
 962                 if (write)
 963                         size = min_t(u64, size, fsc->mount_options->wsize);
 964                 else
 965                         size = min_t(u64, size, fsc->mount_options->rsize);
 966
 967                 vino = ceph_vino(inode);
 968                 req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
 969                                             vino, pos, &size, 0,
 970                                             1,
 971                                             write ? CEPH_OSD_OP_WRITE :
 972                                                     CEPH_OSD_OP_READ,
 973                                             flags, snapc,
 974                                             ci->i_truncate_seq,
 975                                             ci->i_truncate_size,
 976                                             false);
 977                 if (IS_ERR(req)) {
 978                         ret = PTR_ERR(req);
 979                         break;
 980                 }
 981
 982                 len = iter_get_bvecs_alloc(iter, size, &bvecs, &num_pages);
 983                 if (len < 0) {
 984                         ceph_osdc_put_request(req);
 985                         ret = len;
 986                         break;
 987                 }
 988                 if (len != size)
 989                         osd_req_op_extent_update(req, 0, len);
 990
 991                 /*
 992                  * To simplify error handling, allow AIO when IO within i_size
 993                  * or IO can be satisfied by single OSD request.
 994                  */
 995                 if (pos == iocb->ki_pos && !is_sync_kiocb(iocb) &&
 996                     (len == count || pos + count <= i_size_read(inode))) {
 997                         aio_req = kzalloc(sizeof(*aio_req), GFP_KERNEL);
 998                         if (aio_req) {
 999                                 aio_req->iocb = iocb;
1000                                 aio_req->write = write;
1001                                 aio_req->should_dirty = should_dirty;
1002                                 INIT_LIST_HEAD(&aio_req->osd_reqs);
1003                                 if (write) {
1004                                         aio_req->mtime = mtime;
1005                                         swap(aio_req->prealloc_cf, *pcf);
1006                                 }
1007                         }
1008                         /* ignore error */
1009                 }
1010
1011                 if (write) {
1012                         /*
1013                          * throw out any page cache pages in this range. this
1014                          * may block.
1015                          */
1016                         truncate_inode_pages_range(inode->i_mapping, pos,
1017                                                    PAGE_ALIGN(pos + len) - 1);
1018
1019                         req->r_mtime = mtime;
1020                 }
1021
1022                 osd_req_op_extent_osd_data_bvecs(req, 0, bvecs, num_pages, len);
1023
1024                 if (aio_req) {
1025                         aio_req->total_len += len;
1026                         aio_req->num_reqs++;
1027                         atomic_inc(&aio_req->pending_reqs);
1028
1029                         req->r_callback = ceph_aio_complete_req;
1030                         req->r_inode = inode;
1031                         req->r_priv = aio_req;
1032                         list_add_tail(&req->r_private_item, &aio_req->osd_reqs);
1033
1034                         pos += len;
1035                         continue;
1036                 }
1037
1038                 ret = ceph_osdc_start_request(req->r_osdc, req, false);
1039                 if (!ret)
1040                         ret = ceph_osdc_wait_request(&fsc->client->osdc, req);
1041
1042                 size = i_size_read(inode);
1043                 if (!write) {
1044                         if (ret == -ENOENT)
1045                                 ret = 0;
1046                         if (ret >= 0 && ret < len && pos + ret < size) {
1047                                 struct iov_iter i;
1048                                 int zlen = min_t(size_t, len - ret,
1049                                                  size - pos - ret);
1050
1051                                 iov_iter_bvec(&i, READ, bvecs, num_pages, len);
1052                                 iov_iter_advance(&i, ret);
1053                                 iov_iter_zero(zlen, &i);
1054                                 ret += zlen;
1055                         }
1056                         if (ret >= 0)
1057                                 len = ret;
1058                 }
1059
1060                 put_bvecs(bvecs, num_pages, should_dirty);
1061                 ceph_osdc_put_request(req);
1062                 if (ret < 0)
1063                         break;
1064
1065                 pos += len;
1066                 if (!write && pos >= size)
1067                         break;
1068
1069                 if (write && pos > size) {
1070                         if (ceph_inode_set_size(inode, pos))
1071                                 ceph_check_caps(ceph_inode(inode),
1072                                                 CHECK_CAPS_AUTHONLY,
1073                                                 NULL);
1074                 }
1075         }
1076
1077         if (aio_req) {
1078                 LIST_HEAD(osd_reqs);
1079
1080                 if (aio_req->num_reqs == 0) {
1081                         kfree(aio_req);
1082                         return ret;
1083                 }
1084
1085                 ceph_get_cap_refs(ci, write ? CEPH_CAP_FILE_WR :
1086                                               CEPH_CAP_FILE_RD);
1087
1088                 list_splice(&aio_req->osd_reqs, &osd_reqs);
1089                 while (!list_empty(&osd_reqs)) {
1090                         req = list_first_entry(&osd_reqs,
1091                                                struct ceph_osd_request,
1092                                                r_private_item);
1093                         list_del_init(&req->r_private_item);
1094                         if (ret >= 0)
1095                                 ret = ceph_osdc_start_request(req->r_osdc,
1096                                                               req, false);
1097                         if (ret < 0) {
1098                                 req->r_result = ret;
1099                                 ceph_aio_complete_req(req);
1100                         }
1101                 }
1102                 return -EIOCBQUEUED;
1103         }
1104
1105         if (ret != -EOLDSNAPC && pos > iocb->ki_pos) {
1106                 ret = pos - iocb->ki_pos;
1107                 iocb->ki_pos = pos;
1108         }
1109         return ret;
1110 }
1111
1112 /*
1113  * Synchronous write, straight from __user pointer or user pages.
1114  *
1115  * If write spans object boundary, just do multiple writes.  (For a
1116  * correct atomic write, we should e.g. take write locks on all
1117  * objects, rollback on failure, etc.)
1118  */
1119 static ssize_t
1120 ceph_sync_write(struct kiocb *iocb, struct iov_iter *from, loff_t pos,
1121                 struct ceph_snap_context *snapc)
1122 {
1123         struct file *file = iocb->ki_filp;
1124         struct inode *inode = file_inode(file);
1125         struct ceph_inode_info *ci = ceph_inode(inode);
1126         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
1127         struct ceph_vino vino;
1128         struct ceph_osd_request *req;
1129         struct page **pages;
1130         u64 len;
1131         int num_pages;
1132         int written = 0;
1133         int flags;
1134         int ret;
1135         bool check_caps = false;
1136         struct timespec64 mtime = current_time(inode);
1137         size_t count = iov_iter_count(from);
1138
1139         if (ceph_snap(file_inode(file)) != CEPH_NOSNAP)
1140                 return -EROFS;
1141
1142         dout("sync_write on file %p %lld~%u snapc %p seq %lld\n",
1143              file, pos, (unsigned)count, snapc, snapc->seq);
1144
1145         ret = filemap_write_and_wait_range(inode->i_mapping,
1146                                            pos, pos + count - 1);
1147         if (ret < 0)
1148                 return ret;
1149
1150         ret = invalidate_inode_pages2_range(inode->i_mapping,
1151                                             pos >> PAGE_SHIFT,
1152                                             (pos + count - 1) >> PAGE_SHIFT);
1153         if (ret < 0)
1154                 dout("invalidate_inode_pages2_range returned %d\n", ret);
1155
1156         flags = /* CEPH_OSD_FLAG_ORDERSNAP | */ CEPH_OSD_FLAG_WRITE;
1157
1158         while ((len = iov_iter_count(from)) > 0) {
1159                 size_t left;
1160                 int n;
1161
1162                 vino = ceph_vino(inode);
1163                 req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
1164                                             vino, pos, &len, 0, 1,
1165                                             CEPH_OSD_OP_WRITE, flags, snapc,
1166                                             ci->i_truncate_seq,
1167                                             ci->i_truncate_size,
1168                                             false);
1169                 if (IS_ERR(req)) {
1170                         ret = PTR_ERR(req);
1171                         break;
1172                 }
1173
1174                 /*
1175                  * write from beginning of first page,
1176                  * regardless of io alignment
1177                  */
1178                 num_pages = (len + PAGE_SIZE - 1) >> PAGE_SHIFT;
1179
1180                 pages = ceph_alloc_page_vector(num_pages, GFP_KERNEL);
1181                 if (IS_ERR(pages)) {
1182                         ret = PTR_ERR(pages);
1183                         goto out;
1184                 }
1185
1186                 left = len;
1187                 for (n = 0; n < num_pages; n++) {
1188                         size_t plen = min_t(size_t, left, PAGE_SIZE);
1189                         ret = copy_page_from_iter(pages[n], 0, plen, from);
1190                         if (ret != plen) {
1191                                 ret = -EFAULT;
1192                                 break;
1193                         }
1194                         left -= ret;
1195                 }
1196
1197                 if (ret < 0) {
1198                         ceph_release_page_vector(pages, num_pages);
1199                         goto out;
1200                 }
1201
1202                 req->r_inode = inode;
1203
1204                 osd_req_op_extent_osd_data_pages(req, 0, pages, len, 0,
1205                                                 false, true);
1206
1207                 req->r_mtime = mtime;
1208                 ret = ceph_osdc_start_request(&fsc->client->osdc, req, false);
1209                 if (!ret)
1210                         ret = ceph_osdc_wait_request(&fsc->client->osdc, req);
1211
1212 out:
1213                 ceph_osdc_put_request(req);
1214                 if (ret != 0) {
1215                         ceph_set_error_write(ci);
1216                         break;
1217                 }
1218
1219                 ceph_clear_error_write(ci);
1220                 pos += len;
1221                 written += len;
1222                 if (pos > i_size_read(inode)) {
1223                         check_caps = ceph_inode_set_size(inode, pos);
1224                         if (check_caps)
1225                                 ceph_check_caps(ceph_inode(inode),
1226                                                 CHECK_CAPS_AUTHONLY,
1227                                                 NULL);
1228                 }
1229
1230         }
1231
1232         if (ret != -EOLDSNAPC && written > 0) {
1233                 ret = written;
1234                 iocb->ki_pos = pos;
1235         }
1236         return ret;
1237 }
1238
1239 /*
1240  * Wrap generic_file_aio_read with checks for cap bits on the inode.
1241  * Atomically grab references, so that those bits are not released
1242  * back to the MDS mid-read.
1243  *
1244  * Hmm, the sync read case isn't actually async... should it be?
1245  */
1246 static ssize_t ceph_read_iter(struct kiocb *iocb, struct iov_iter *to)
1247 {
1248         struct file *filp = iocb->ki_filp;
1249         struct ceph_file_info *fi = filp->private_data;
1250         size_t len = iov_iter_count(to);
1251         struct inode *inode = file_inode(filp);
1252         struct ceph_inode_info *ci = ceph_inode(inode);
1253         struct page *pinned_page = NULL;
1254         ssize_t ret;
1255         int want, got = 0;
1256         int retry_op = 0, read = 0;
1257
1258 again:
1259         dout("aio_read %p %llx.%llx %llu~%u trying to get caps on %p\n",
1260              inode, ceph_vinop(inode), iocb->ki_pos, (unsigned)len, inode);
1261
1262         if (fi->fmode & CEPH_FILE_MODE_LAZY)
1263                 want = CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_LAZYIO;
1264         else
1265                 want = CEPH_CAP_FILE_CACHE;
1266         ret = ceph_get_caps(ci, CEPH_CAP_FILE_RD, want, -1, &got, &pinned_page);
1267         if (ret < 0)
1268                 return ret;
1269
1270         if ((got & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) == 0 ||
1271             (iocb->ki_flags & IOCB_DIRECT) ||
1272             (fi->flags & CEPH_F_SYNC)) {
1273
1274                 dout("aio_sync_read %p %llx.%llx %llu~%u got cap refs on %s\n",
1275                      inode, ceph_vinop(inode), iocb->ki_pos, (unsigned)len,
1276                      ceph_cap_string(got));
1277
1278                 if (ci->i_inline_version == CEPH_INLINE_NONE) {
1279                         if (!retry_op && (iocb->ki_flags & IOCB_DIRECT)) {
1280                                 ret = ceph_direct_read_write(iocb, to,
1281                                                              NULL, NULL);
1282                                 if (ret >= 0 && ret < len)
1283                                         retry_op = CHECK_EOF;
1284                         } else {
1285                                 ret = ceph_sync_read(iocb, to, &retry_op);
1286                         }
1287                 } else {
1288                         retry_op = READ_INLINE;
1289                 }
1290         } else {
1291                 CEPH_DEFINE_RW_CONTEXT(rw_ctx, got);
1292                 dout("aio_read %p %llx.%llx %llu~%u got cap refs on %s\n",
1293                      inode, ceph_vinop(inode), iocb->ki_pos, (unsigned)len,
1294                      ceph_cap_string(got));
1295                 ceph_add_rw_context(fi, &rw_ctx);
1296                 ret = generic_file_read_iter(iocb, to);
1297                 ceph_del_rw_context(fi, &rw_ctx);
1298         }
1299         dout("aio_read %p %llx.%llx dropping cap refs on %s = %d\n",
1300              inode, ceph_vinop(inode), ceph_cap_string(got), (int)ret);
1301         if (pinned_page) {
1302                 put_page(pinned_page);
1303                 pinned_page = NULL;
1304         }
1305         ceph_put_cap_refs(ci, got);
1306         if (retry_op > HAVE_RETRIED && ret >= 0) {
1307                 int statret;
1308                 struct page *page = NULL;
1309                 loff_t i_size;
1310                 if (retry_op == READ_INLINE) {
1311                         page = __page_cache_alloc(GFP_KERNEL);
1312                         if (!page)
1313                                 return -ENOMEM;
1314                 }
1315
1316                 statret = __ceph_do_getattr(inode, page,
1317                                             CEPH_STAT_CAP_INLINE_DATA, !!page);
1318                 if (statret < 0) {
1319                         if (page)
1320                                 __free_page(page);
1321                         if (statret == -ENODATA) {
1322                                 BUG_ON(retry_op != READ_INLINE);
1323                                 goto again;
1324                         }
1325                         return statret;
1326                 }
1327
1328                 i_size = i_size_read(inode);
1329                 if (retry_op == READ_INLINE) {
1330                         BUG_ON(ret > 0 || read > 0);
1331                         if (iocb->ki_pos < i_size &&
1332                             iocb->ki_pos < PAGE_SIZE) {
1333                                 loff_t end = min_t(loff_t, i_size,
1334                                                    iocb->ki_pos + len);
1335                                 end = min_t(loff_t, end, PAGE_SIZE);
1336                                 if (statret < end)
1337                                         zero_user_segment(page, statret, end);
1338                                 ret = copy_page_to_iter(page,
1339                                                 iocb->ki_pos & ~PAGE_MASK,
1340                                                 end - iocb->ki_pos, to);
1341                                 iocb->ki_pos += ret;
1342                                 read += ret;
1343                         }
1344                         if (iocb->ki_pos < i_size && read < len) {
1345                                 size_t zlen = min_t(size_t, len - read,
1346                                                     i_size - iocb->ki_pos);
1347                                 ret = iov_iter_zero(zlen, to);
1348                                 iocb->ki_pos += ret;
1349                                 read += ret;
1350                         }
1351                         __free_pages(page, 0);
1352                         return read;
1353                 }
1354
1355                 /* hit EOF or hole? */
1356                 if (retry_op == CHECK_EOF && iocb->ki_pos < i_size &&
1357                     ret < len) {
1358                         dout("sync_read hit hole, ppos %lld < size %lld"
1359                              ", reading more\n", iocb->ki_pos, i_size);
1360
1361                         read += ret;
1362                         len -= ret;
1363                         retry_op = HAVE_RETRIED;
1364                         goto again;
1365                 }
1366         }
1367
1368         if (ret >= 0)
1369                 ret += read;
1370
1371         return ret;
1372 }
1373
1374 /*
1375  * Take cap references to avoid releasing caps to MDS mid-write.
1376  *
1377  * If we are synchronous, and write with an old snap context, the OSD
1378  * may return EOLDSNAPC.  In that case, retry the write.. _after_
1379  * dropping our cap refs and allowing the pending snap to logically
1380  * complete _before_ this write occurs.
1381  *
1382  * If we are near ENOSPC, write synchronously.
1383  */
1384 static ssize_t ceph_write_iter(struct kiocb *iocb, struct iov_iter *from)
1385 {
1386         struct file *file = iocb->ki_filp;
1387         struct ceph_file_info *fi = file->private_data;
1388         struct inode *inode = file_inode(file);
1389         struct ceph_inode_info *ci = ceph_inode(inode);
1390         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
1391         struct ceph_cap_flush *prealloc_cf;
1392         ssize_t count, written = 0;
1393         int err, want, got;
1394         loff_t pos;
1395         loff_t limit = max(i_size_read(inode), fsc->max_file_size);
1396
1397         if (ceph_snap(inode) != CEPH_NOSNAP)
1398                 return -EROFS;
1399
1400         prealloc_cf = ceph_alloc_cap_flush();
1401         if (!prealloc_cf)
1402                 return -ENOMEM;
1403
1404 retry_snap:
1405         inode_lock(inode);
1406
1407         /* We can write back this queue in page reclaim */
1408         current->backing_dev_info = inode_to_bdi(inode);
1409
1410         if (iocb->ki_flags & IOCB_APPEND) {
1411                 err = ceph_do_getattr(inode, CEPH_STAT_CAP_SIZE, false);
1412                 if (err < 0)
1413                         goto out;
1414         }
1415
1416         err = generic_write_checks(iocb, from);
1417         if (err <= 0)
1418                 goto out;
1419
1420         pos = iocb->ki_pos;
1421         if (unlikely(pos >= limit)) {
1422                 err = -EFBIG;
1423                 goto out;
1424         } else {
1425                 iov_iter_truncate(from, limit - pos);
1426         }
1427
1428         count = iov_iter_count(from);
1429         if (ceph_quota_is_max_bytes_exceeded(inode, pos + count)) {
1430                 err = -EDQUOT;
1431                 goto out;
1432         }
1433
1434         err = file_remove_privs(file);
1435         if (err)
1436                 goto out;
1437
1438         err = file_update_time(file);
1439         if (err)
1440                 goto out;
1441
1442         inode_inc_iversion_raw(inode);
1443
1444         if (ci->i_inline_version != CEPH_INLINE_NONE) {
1445                 err = ceph_uninline_data(file, NULL);
1446                 if (err < 0)
1447                         goto out;
1448         }
1449
1450         /* FIXME: not complete since it doesn't account for being at quota */
1451         if (ceph_osdmap_flag(&fsc->client->osdc, CEPH_OSDMAP_FULL)) {
1452                 err = -ENOSPC;
1453                 goto out;
1454         }
1455
1456         dout("aio_write %p %llx.%llx %llu~%zd getting caps. i_size %llu\n",
1457              inode, ceph_vinop(inode), pos, count, i_size_read(inode));
1458         if (fi->fmode & CEPH_FILE_MODE_LAZY)
1459                 want = CEPH_CAP_FILE_BUFFER | CEPH_CAP_FILE_LAZYIO;
1460         else
1461                 want = CEPH_CAP_FILE_BUFFER;
1462         got = 0;
1463         err = ceph_get_caps(ci, CEPH_CAP_FILE_WR, want, pos + count,
1464                             &got, NULL);
1465         if (err < 0)
1466                 goto out;
1467
1468         dout("aio_write %p %llx.%llx %llu~%zd got cap refs on %s\n",
1469              inode, ceph_vinop(inode), pos, count, ceph_cap_string(got));
1470
1471         if ((got & (CEPH_CAP_FILE_BUFFER|CEPH_CAP_FILE_LAZYIO)) == 0 ||
1472             (iocb->ki_flags & IOCB_DIRECT) || (fi->flags & CEPH_F_SYNC) ||
1473             (ci->i_ceph_flags & CEPH_I_ERROR_WRITE)) {
1474                 struct ceph_snap_context *snapc;
1475                 struct iov_iter data;
1476                 inode_unlock(inode);
1477
1478                 spin_lock(&ci->i_ceph_lock);
1479                 if (__ceph_have_pending_cap_snap(ci)) {
1480                         struct ceph_cap_snap *capsnap =
1481                                         list_last_entry(&ci->i_cap_snaps,
1482                                                         struct ceph_cap_snap,
1483                                                         ci_item);
1484                         snapc = ceph_get_snap_context(capsnap->context);
1485                 } else {
1486                         BUG_ON(!ci->i_head_snapc);
1487                         snapc = ceph_get_snap_context(ci->i_head_snapc);
1488                 }
1489                 spin_unlock(&ci->i_ceph_lock);
1490
1491                 /* we might need to revert back to that point */
1492                 data = *from;
1493                 if (iocb->ki_flags & IOCB_DIRECT)
1494                         written = ceph_direct_read_write(iocb, &data, snapc,
1495                                                          &prealloc_cf);
1496                 else
1497                         written = ceph_sync_write(iocb, &data, pos, snapc);
1498                 if (written > 0)
1499                         iov_iter_advance(from, written);
1500                 ceph_put_snap_context(snapc);
1501         } else {
1502                 /*
1503                  * No need to acquire the i_truncate_mutex. Because
1504                  * the MDS revokes Fwb caps before sending truncate
1505                  * message to us. We can't get Fwb cap while there
1506                  * are pending vmtruncate. So write and vmtruncate
1507                  * can not run at the same time
1508                  */
1509                 written = generic_perform_write(file, from, pos);
1510                 if (likely(written >= 0))
1511                         iocb->ki_pos = pos + written;
1512                 inode_unlock(inode);
1513         }
1514
1515         if (written >= 0) {
1516                 int dirty;
1517
1518                 spin_lock(&ci->i_ceph_lock);
1519                 ci->i_inline_version = CEPH_INLINE_NONE;
1520                 dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR,
1521                                                &prealloc_cf);
1522                 spin_unlock(&ci->i_ceph_lock);
1523                 if (dirty)
1524                         __mark_inode_dirty(inode, dirty);
1525                 if (ceph_quota_is_max_bytes_approaching(inode, iocb->ki_pos))
1526                         ceph_check_caps(ci, CHECK_CAPS_NODELAY, NULL);
1527         }
1528
1529         dout("aio_write %p %llx.%llx %llu~%u  dropping cap refs on %s\n",
1530              inode, ceph_vinop(inode), pos, (unsigned)count,
1531              ceph_cap_string(got));
1532         ceph_put_cap_refs(ci, got);
1533
1534         if (written == -EOLDSNAPC) {
1535                 dout("aio_write %p %llx.%llx %llu~%u" "got EOLDSNAPC, retrying\n",
1536                      inode, ceph_vinop(inode), pos, (unsigned)count);
1537                 goto retry_snap;
1538         }
1539
1540         if (written >= 0) {
1541                 if (ceph_osdmap_flag(&fsc->client->osdc, CEPH_OSDMAP_NEARFULL))
1542                         iocb->ki_flags |= IOCB_DSYNC;
1543                 written = generic_write_sync(iocb, written);
1544         }
1545
1546         goto out_unlocked;
1547
1548 out:
1549         inode_unlock(inode);
1550 out_unlocked:
1551         ceph_free_cap_flush(prealloc_cf);
1552         current->backing_dev_info = NULL;
1553         return written ? written : err;
1554 }
1555
1556 /*
1557  * llseek.  be sure to verify file size on SEEK_END.
1558  */
1559 static loff_t ceph_llseek(struct file *file, loff_t offset, int whence)
1560 {
1561         struct inode *inode = file->f_mapping->host;
1562         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
1563         loff_t i_size;
1564         loff_t ret;
1565
1566         inode_lock(inode);
1567
1568         if (whence == SEEK_END || whence == SEEK_DATA || whence == SEEK_HOLE) {
1569                 ret = ceph_do_getattr(inode, CEPH_STAT_CAP_SIZE, false);
1570                 if (ret < 0)
1571                         goto out;
1572         }
1573
1574         i_size = i_size_read(inode);
1575         switch (whence) {
1576         case SEEK_END:
1577                 offset += i_size;
1578                 break;
1579         case SEEK_CUR:
1580                 /*
1581                  * Here we special-case the lseek(fd, 0, SEEK_CUR)
1582                  * position-querying operation.  Avoid rewriting the "same"
1583                  * f_pos value back to the file because a concurrent read(),
1584                  * write() or lseek() might have altered it
1585                  */
1586                 if (offset == 0) {
1587                         ret = file->f_pos;
1588                         goto out;
1589                 }
1590                 offset += file->f_pos;
1591                 break;
1592         case SEEK_DATA:
1593                 if (offset < 0 || offset >= i_size) {
1594                         ret = -ENXIO;
1595                         goto out;
1596                 }
1597                 break;
1598         case SEEK_HOLE:
1599                 if (offset < 0 || offset >= i_size) {
1600                         ret = -ENXIO;
1601                         goto out;
1602                 }
1603                 offset = i_size;
1604                 break;
1605         }
1606
1607         ret = vfs_setpos(file, offset, max(i_size, fsc->max_file_size));
1608
1609 out:
1610         inode_unlock(inode);
1611         return ret;
1612 }
1613
1614 static inline void ceph_zero_partial_page(
1615         struct inode *inode, loff_t offset, unsigned size)
1616 {
1617         struct page *page;
1618         pgoff_t index = offset >> PAGE_SHIFT;
1619
1620         page = find_lock_page(inode->i_mapping, index);
1621         if (page) {
1622                 wait_on_page_writeback(page);
1623                 zero_user(page, offset & (PAGE_SIZE - 1), size);
1624                 unlock_page(page);
1625                 put_page(page);
1626         }
1627 }
1628
1629 static void ceph_zero_pagecache_range(struct inode *inode, loff_t offset,
1630                                       loff_t length)
1631 {
1632         loff_t nearly = round_up(offset, PAGE_SIZE);
1633         if (offset < nearly) {
1634                 loff_t size = nearly - offset;
1635                 if (length < size)
1636                         size = length;
1637                 ceph_zero_partial_page(inode, offset, size);
1638                 offset += size;
1639                 length -= size;
1640         }
1641         if (length >= PAGE_SIZE) {
1642                 loff_t size = round_down(length, PAGE_SIZE);
1643                 truncate_pagecache_range(inode, offset, offset + size - 1);
1644                 offset += size;
1645                 length -= size;
1646         }
1647         if (length)
1648                 ceph_zero_partial_page(inode, offset, length);
1649 }
1650
1651 static int ceph_zero_partial_object(struct inode *inode,
1652                                     loff_t offset, loff_t *length)
1653 {
1654         struct ceph_inode_info *ci = ceph_inode(inode);
1655         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
1656         struct ceph_osd_request *req;
1657         int ret = 0;
1658         loff_t zero = 0;
1659         int op;
1660
1661         if (!length) {
1662                 op = offset ? CEPH_OSD_OP_DELETE : CEPH_OSD_OP_TRUNCATE;
1663                 length = &zero;
1664         } else {
1665                 op = CEPH_OSD_OP_ZERO;
1666         }
1667
1668         req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
1669                                         ceph_vino(inode),
1670                                         offset, length,
1671                                         0, 1, op,
1672                                         CEPH_OSD_FLAG_WRITE,
1673                                         NULL, 0, 0, false);
1674         if (IS_ERR(req)) {
1675                 ret = PTR_ERR(req);
1676                 goto out;
1677         }
1678
1679         req->r_mtime = inode->i_mtime;
1680         ret = ceph_osdc_start_request(&fsc->client->osdc, req, false);
1681         if (!ret) {
1682                 ret = ceph_osdc_wait_request(&fsc->client->osdc, req);
1683                 if (ret == -ENOENT)
1684                         ret = 0;
1685         }
1686         ceph_osdc_put_request(req);
1687
1688 out:
1689         return ret;
1690 }
1691
1692 static int ceph_zero_objects(struct inode *inode, loff_t offset, loff_t length)
1693 {
1694         int ret = 0;
1695         struct ceph_inode_info *ci = ceph_inode(inode);
1696         s32 stripe_unit = ci->i_layout.stripe_unit;
1697         s32 stripe_count = ci->i_layout.stripe_count;
1698         s32 object_size = ci->i_layout.object_size;
1699         u64 object_set_size = object_size * stripe_count;
1700         u64 nearly, t;
1701
1702         /* round offset up to next period boundary */
1703         nearly = offset + object_set_size - 1;
1704         t = nearly;
1705         nearly -= do_div(t, object_set_size);
1706
1707         while (length && offset < nearly) {
1708                 loff_t size = length;
1709                 ret = ceph_zero_partial_object(inode, offset, &size);
1710                 if (ret < 0)
1711                         return ret;
1712                 offset += size;
1713                 length -= size;
1714         }
1715         while (length >= object_set_size) {
1716                 int i;
1717                 loff_t pos = offset;
1718                 for (i = 0; i < stripe_count; ++i) {
1719                         ret = ceph_zero_partial_object(inode, pos, NULL);
1720                         if (ret < 0)
1721                                 return ret;
1722                         pos += stripe_unit;
1723                 }
1724                 offset += object_set_size;
1725                 length -= object_set_size;
1726         }
1727         while (length) {
1728                 loff_t size = length;
1729                 ret = ceph_zero_partial_object(inode, offset, &size);
1730                 if (ret < 0)
1731                         return ret;
1732                 offset += size;
1733                 length -= size;
1734         }
1735         return ret;
1736 }
1737
1738 static long ceph_fallocate(struct file *file, int mode,
1739                                 loff_t offset, loff_t length)
1740 {
1741         struct ceph_file_info *fi = file->private_data;
1742         struct inode *inode = file_inode(file);
1743         struct ceph_inode_info *ci = ceph_inode(inode);
1744         struct ceph_cap_flush *prealloc_cf;
1745         int want, got = 0;
1746         int dirty;
1747         int ret = 0;
1748         loff_t endoff = 0;
1749         loff_t size;
1750
1751         if (mode != (FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
1752                 return -EOPNOTSUPP;
1753
1754         if (!S_ISREG(inode->i_mode))
1755                 return -EOPNOTSUPP;
1756
1757         prealloc_cf = ceph_alloc_cap_flush();
1758         if (!prealloc_cf)
1759                 return -ENOMEM;
1760
1761         inode_lock(inode);
1762
1763         if (ceph_snap(inode) != CEPH_NOSNAP) {
1764                 ret = -EROFS;
1765                 goto unlock;
1766         }
1767
1768         if (ci->i_inline_version != CEPH_INLINE_NONE) {
1769                 ret = ceph_uninline_data(file, NULL);
1770                 if (ret < 0)
1771                         goto unlock;
1772         }
1773
1774         size = i_size_read(inode);
1775
1776         /* Are we punching a hole beyond EOF? */
1777         if (offset >= size)
1778                 goto unlock;
1779         if ((offset + length) > size)
1780                 length = size - offset;
1781
1782         if (fi->fmode & CEPH_FILE_MODE_LAZY)
1783                 want = CEPH_CAP_FILE_BUFFER | CEPH_CAP_FILE_LAZYIO;
1784         else
1785                 want = CEPH_CAP_FILE_BUFFER;
1786
1787         ret = ceph_get_caps(ci, CEPH_CAP_FILE_WR, want, endoff, &got, NULL);
1788         if (ret < 0)
1789                 goto unlock;
1790
1791         ceph_zero_pagecache_range(inode, offset, length);
1792         ret = ceph_zero_objects(inode, offset, length);
1793
1794         if (!ret) {
1795                 spin_lock(&ci->i_ceph_lock);
1796                 ci->i_inline_version = CEPH_INLINE_NONE;
1797                 dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR,
1798                                                &prealloc_cf);
1799                 spin_unlock(&ci->i_ceph_lock);
1800                 if (dirty)
1801                         __mark_inode_dirty(inode, dirty);
1802         }
1803
1804         ceph_put_cap_refs(ci, got);
1805 unlock:
1806         inode_unlock(inode);
1807         ceph_free_cap_flush(prealloc_cf);
1808         return ret;
1809 }
1810
1811 /*
1812  * This function tries to get FILE_WR capabilities for dst_ci and FILE_RD for
1813  * src_ci.  Two attempts are made to obtain both caps, and an error is return if
1814  * this fails; zero is returned on success.
1815  */
1816 static int get_rd_wr_caps(struct ceph_inode_info *src_ci,
1817                           loff_t src_endoff, int *src_got,
1818                           struct ceph_inode_info *dst_ci,
1819                           loff_t dst_endoff, int *dst_got)
1820 {
1821         int ret = 0;
1822         bool retrying = false;
1823
1824 retry_caps:
1825         ret = ceph_get_caps(dst_ci, CEPH_CAP_FILE_WR, CEPH_CAP_FILE_BUFFER,
1826                             dst_endoff, dst_got, NULL);
1827         if (ret < 0)
1828                 return ret;
1829
1830         /*
1831          * Since we're already holding the FILE_WR capability for the dst file,
1832          * we would risk a deadlock by using ceph_get_caps.  Thus, we'll do some
1833          * retry dance instead to try to get both capabilities.
1834          */
1835         ret = ceph_try_get_caps(src_ci, CEPH_CAP_FILE_RD, CEPH_CAP_FILE_SHARED,
1836                                 false, src_got);
1837         if (ret <= 0) {
1838                 /* Start by dropping dst_ci caps and getting src_ci caps */
1839                 ceph_put_cap_refs(dst_ci, *dst_got);
1840                 if (retrying) {
1841                         if (!ret)
1842                                 /* ceph_try_get_caps masks EAGAIN */
1843                                 ret = -EAGAIN;
1844                         return ret;
1845                 }
1846                 ret = ceph_get_caps(src_ci, CEPH_CAP_FILE_RD,
1847                                     CEPH_CAP_FILE_SHARED, src_endoff,
1848                                     src_got, NULL);
1849                 if (ret < 0)
1850                         return ret;
1851                 /*... drop src_ci caps too, and retry */
1852                 ceph_put_cap_refs(src_ci, *src_got);
1853                 retrying = true;
1854                 goto retry_caps;
1855         }
1856         return ret;
1857 }
1858
1859 static void put_rd_wr_caps(struct ceph_inode_info *src_ci, int src_got,
1860                            struct ceph_inode_info *dst_ci, int dst_got)
1861 {
1862         ceph_put_cap_refs(src_ci, src_got);
1863         ceph_put_cap_refs(dst_ci, dst_got);
1864 }
1865
1866 /*
1867  * This function does several size-related checks, returning an error if:
1868  *  - source file is smaller than off+len
1869  *  - destination file size is not OK (inode_newsize_ok())
1870  *  - max bytes quotas is exceeded
1871  */
1872 static int is_file_size_ok(struct inode *src_inode, struct inode *dst_inode,
1873                            loff_t src_off, loff_t dst_off, size_t len)
1874 {
1875         loff_t size, endoff;
1876
1877         size = i_size_read(src_inode);
1878         /*
1879          * Don't copy beyond source file EOF.  Instead of simply setting length
1880          * to (size - src_off), just drop to VFS default implementation, as the
1881          * local i_size may be stale due to other clients writing to the source
1882          * inode.
1883          */
1884         if (src_off + len > size) {
1885                 dout("Copy beyond EOF (%llu + %zu > %llu)\n",
1886                      src_off, len, size);
1887                 return -EOPNOTSUPP;
1888         }
1889         size = i_size_read(dst_inode);
1890
1891         endoff = dst_off + len;
1892         if (inode_newsize_ok(dst_inode, endoff))
1893                 return -EOPNOTSUPP;
1894
1895         if (ceph_quota_is_max_bytes_exceeded(dst_inode, endoff))
1896                 return -EDQUOT;
1897
1898         return 0;
1899 }
1900
1901 static ssize_t __ceph_copy_file_range(struct file *src_file, loff_t src_off,
1902                                       struct file *dst_file, loff_t dst_off,
1903                                       size_t len, unsigned int flags)
1904 {
1905         struct inode *src_inode = file_inode(src_file);
1906         struct inode *dst_inode = file_inode(dst_file);
1907         struct ceph_inode_info *src_ci = ceph_inode(src_inode);
1908         struct ceph_inode_info *dst_ci = ceph_inode(dst_inode);
1909         struct ceph_cap_flush *prealloc_cf;
1910         struct ceph_object_locator src_oloc, dst_oloc;
1911         struct ceph_object_id src_oid, dst_oid;
1912         loff_t endoff = 0, size;
1913         ssize_t ret = -EIO;
1914         u64 src_objnum, dst_objnum, src_objoff, dst_objoff;
1915         u32 src_objlen, dst_objlen, object_size;
1916         int src_got = 0, dst_got = 0, err, dirty;
1917         bool do_final_copy = false;
1918
1919         if (src_inode == dst_inode)
1920                 return -EINVAL;
1921         if (src_inode->i_sb != dst_inode->i_sb)
1922                 return -EXDEV;
1923         if (ceph_snap(dst_inode) != CEPH_NOSNAP)
1924                 return -EROFS;
1925
1926         /*
1927          * Some of the checks below will return -EOPNOTSUPP, which will force a
1928          * fallback to the default VFS copy_file_range implementation.  This is
1929          * desirable in several cases (for ex, the 'len' is smaller than the
1930          * size of the objects, or in cases where that would be more
1931          * efficient).
1932          */
1933
1934         if (ceph_test_mount_opt(ceph_inode_to_client(src_inode), NOCOPYFROM))
1935                 return -EOPNOTSUPP;
1936
1937         /*
1938          * Striped file layouts require that we copy partial objects, but the
1939          * OSD copy-from operation only supports full-object copies.  Limit
1940          * this to non-striped file layouts for now.
1941          */
1942         if ((src_ci->i_layout.stripe_unit != dst_ci->i_layout.stripe_unit) ||
1943             (src_ci->i_layout.stripe_count != 1) ||
1944             (dst_ci->i_layout.stripe_count != 1) ||
1945             (src_ci->i_layout.object_size != dst_ci->i_layout.object_size)) {
1946                 dout("Invalid src/dst files layout\n");
1947                 return -EOPNOTSUPP;
1948         }
1949
1950         if (len < src_ci->i_layout.object_size)
1951                 return -EOPNOTSUPP; /* no remote copy will be done */
1952
1953         prealloc_cf = ceph_alloc_cap_flush();
1954         if (!prealloc_cf)
1955                 return -ENOMEM;
1956
1957         /* Start by sync'ing the source and destination files */
1958         ret = file_write_and_wait_range(src_file, src_off, (src_off + len));
1959         if (ret < 0) {
1960                 dout("failed to write src file (%zd)\n", ret);
1961                 goto out;
1962         }
1963         ret = file_write_and_wait_range(dst_file, dst_off, (dst_off + len));
1964         if (ret < 0) {
1965                 dout("failed to write dst file (%zd)\n", ret);
1966                 goto out;
1967         }
1968
1969         /*
1970          * We need FILE_WR caps for dst_ci and FILE_RD for src_ci as other
1971          * clients may have dirty data in their caches.  And OSDs know nothing
1972          * about caps, so they can't safely do the remote object copies.
1973          */
1974         err = get_rd_wr_caps(src_ci, (src_off + len), &src_got,
1975                              dst_ci, (dst_off + len), &dst_got);
1976         if (err < 0) {
1977                 dout("get_rd_wr_caps returned %d\n", err);
1978                 ret = -EOPNOTSUPP;
1979                 goto out;
1980         }
1981
1982         ret = is_file_size_ok(src_inode, dst_inode, src_off, dst_off, len);
1983         if (ret < 0)
1984                 goto out_caps;
1985
1986         size = i_size_read(dst_inode);
1987         endoff = dst_off + len;
1988
1989         /* Drop dst file cached pages */
1990         ret = invalidate_inode_pages2_range(dst_inode->i_mapping,
1991                                             dst_off >> PAGE_SHIFT,
1992                                             endoff >> PAGE_SHIFT);
1993         if (ret < 0) {
1994                 dout("Failed to invalidate inode pages (%zd)\n", ret);
1995                 ret = 0; /* XXX */
1996         }
1997         src_oloc.pool = src_ci->i_layout.pool_id;
1998         src_oloc.pool_ns = ceph_try_get_string(src_ci->i_layout.pool_ns);
1999         dst_oloc.pool = dst_ci->i_layout.pool_id;
2000         dst_oloc.pool_ns = ceph_try_get_string(dst_ci->i_layout.pool_ns);
2001
2002         ceph_calc_file_object_mapping(&src_ci->i_layout, src_off,
2003                                       src_ci->i_layout.object_size,
2004                                       &src_objnum, &src_objoff, &src_objlen);
2005         ceph_calc_file_object_mapping(&dst_ci->i_layout, dst_off,
2006                                       dst_ci->i_layout.object_size,
2007                                       &dst_objnum, &dst_objoff, &dst_objlen);
2008         /* object-level offsets need to the same */
2009         if (src_objoff != dst_objoff) {
2010                 ret = -EOPNOTSUPP;
2011                 goto out_caps;
2012         }
2013
2014         /*
2015          * Do a manual copy if the object offset isn't object aligned.
2016          * 'src_objlen' contains the bytes left until the end of the object,
2017          * starting at the src_off
2018          */
2019         if (src_objoff) {
2020                 /*
2021                  * we need to temporarily drop all caps as we'll be calling
2022                  * {read,write}_iter, which will get caps again.
2023                  */
2024                 put_rd_wr_caps(src_ci, src_got, dst_ci, dst_got);
2025                 ret = do_splice_direct(src_file, &src_off, dst_file,
2026                                        &dst_off, src_objlen, flags);
2027                 if (ret < 0) {
2028                         dout("do_splice_direct returned %d\n", err);
2029                         goto out;
2030                 }
2031                 len -= ret;
2032                 err = get_rd_wr_caps(src_ci, (src_off + len),
2033                                      &src_got, dst_ci,
2034                                      (dst_off + len), &dst_got);
2035                 if (err < 0)
2036                         goto out;
2037                 err = is_file_size_ok(src_inode, dst_inode,
2038                                       src_off, dst_off, len);
2039                 if (err < 0)
2040                         goto out_caps;
2041         }
2042         object_size = src_ci->i_layout.object_size;
2043         while (len >= object_size) {
2044                 ceph_calc_file_object_mapping(&src_ci->i_layout, src_off,
2045                                               object_size, &src_objnum,
2046                                               &src_objoff, &src_objlen);
2047                 ceph_calc_file_object_mapping(&dst_ci->i_layout, dst_off,
2048                                               object_size, &dst_objnum,
2049                                               &dst_objoff, &dst_objlen);
2050                 ceph_oid_init(&src_oid);
2051                 ceph_oid_printf(&src_oid, "%llx.%08llx",
2052                                 src_ci->i_vino.ino, src_objnum);
2053                 ceph_oid_init(&dst_oid);
2054                 ceph_oid_printf(&dst_oid, "%llx.%08llx",
2055                                 dst_ci->i_vino.ino, dst_objnum);
2056                 /* Do an object remote copy */
2057                 err = ceph_osdc_copy_from(
2058                         &ceph_inode_to_client(src_inode)->client->osdc,
2059                         src_ci->i_vino.snap, 0,
2060                         &src_oid, &src_oloc,
2061                         CEPH_OSD_OP_FLAG_FADVISE_SEQUENTIAL |
2062                         CEPH_OSD_OP_FLAG_FADVISE_NOCACHE,
2063                         &dst_oid, &dst_oloc,
2064                         CEPH_OSD_OP_FLAG_FADVISE_SEQUENTIAL |
2065                         CEPH_OSD_OP_FLAG_FADVISE_DONTNEED, 0);
2066                 if (err) {
2067                         dout("ceph_osdc_copy_from returned %d\n", err);
2068                         if (!ret)
2069                                 ret = err;
2070                         goto out_caps;
2071                 }
2072                 len -= object_size;
2073                 src_off += object_size;
2074                 dst_off += object_size;
2075                 ret += object_size;
2076         }
2077
2078         if (len)
2079                 /* We still need one final local copy */
2080                 do_final_copy = true;
2081
2082         file_update_time(dst_file);
2083         inode_inc_iversion_raw(dst_inode);
2084
2085         if (endoff > size) {
2086                 int caps_flags = 0;
2087
2088                 /* Let the MDS know about dst file size change */
2089                 if (ceph_quota_is_max_bytes_approaching(dst_inode, endoff))
2090                         caps_flags |= CHECK_CAPS_NODELAY;
2091                 if (ceph_inode_set_size(dst_inode, endoff))
2092                         caps_flags |= CHECK_CAPS_AUTHONLY;
2093                 if (caps_flags)
2094                         ceph_check_caps(dst_ci, caps_flags, NULL);
2095         }
2096         /* Mark Fw dirty */
2097         spin_lock(&dst_ci->i_ceph_lock);
2098         dst_ci->i_inline_version = CEPH_INLINE_NONE;
2099         dirty = __ceph_mark_dirty_caps(dst_ci, CEPH_CAP_FILE_WR, &prealloc_cf);
2100         spin_unlock(&dst_ci->i_ceph_lock);
2101         if (dirty)
2102                 __mark_inode_dirty(dst_inode, dirty);
2103
2104 out_caps:
2105         put_rd_wr_caps(src_ci, src_got, dst_ci, dst_got);
2106
2107         if (do_final_copy) {
2108                 err = do_splice_direct(src_file, &src_off, dst_file,
2109                                        &dst_off, len, flags);
2110                 if (err < 0) {
2111                         dout("do_splice_direct returned %d\n", err);
2112                         goto out;
2113                 }
2114                 len -= err;
2115                 ret += err;
2116         }
2117
2118 out:
2119         ceph_free_cap_flush(prealloc_cf);
2120
2121         return ret;
2122 }
2123
2124 static ssize_t ceph_copy_file_range(struct file *src_file, loff_t src_off,
2125                                     struct file *dst_file, loff_t dst_off,
2126                                     size_t len, unsigned int flags)
2127 {
2128         ssize_t ret;
2129
2130         ret = __ceph_copy_file_range(src_file, src_off, dst_file, dst_off,
2131                                      len, flags);
2132
2133         if (ret == -EOPNOTSUPP || ret == -EXDEV)
2134                 ret = generic_copy_file_range(src_file, src_off, dst_file,
2135                                               dst_off, len, flags);
2136         return ret;
2137 }
2138
2139 const struct file_operations ceph_file_fops = {
2140         .open = ceph_open,
2141         .release = ceph_release,
2142         .llseek = ceph_llseek,
2143         .read_iter = ceph_read_iter,
2144         .write_iter = ceph_write_iter,
2145         .mmap = ceph_mmap,
2146         .fsync = ceph_fsync,
2147         .lock = ceph_lock,
2148         .flock = ceph_flock,
2149         .splice_read = generic_file_splice_read,
2150         .splice_write = iter_file_splice_write,
2151         .unlocked_ioctl = ceph_ioctl,
2152         .compat_ioctl   = ceph_ioctl,
2153         .fallocate      = ceph_fallocate,
2154         .copy_file_range = ceph_copy_file_range,
2155 };